Four new cards from three operators across three lanes. Tom Tunguz on what AI actually does to output quality, measured across 200 pieces. Joseph Kim on two related gaps: the citation graph that governs AI game discovery, and the incentive structure that caps AI ROI in salaried organizations. Simon Willison on system prompts as reactive legal compliance tools, and on the gap between published prompts and actual behavioral specifications.
AI and output quality: floor rises faster than ceiling
Tom Tunguz published a measurement study of his own writing output across 200 pieces after integrating a closed-loop AI taste flywheel. The headline number is counterintuitive: draft churn fell from 47 to 3, but line-level edits stayed constant at a median of 136 per piece. The human effort did not decrease. It relocated. The 10th-percentile quality score rose from 2.59 to 3.81, a larger gain than the 90th percentile saw.
"AI does not shrink human effort. It raises the ceiling of what the same amount of work produces."
The mechanism: AI eliminates structural drafting failures but cannot eliminate taste-driven variation at the top. The floor rises disproportionately because the worst output was previously bottlenecked by the time required to generate a structurally correct draft. AI raises the quality floor more than the ceiling by concentrating human effort on rhetorical refinement.
Joseph Kim's AI ROI piece offers the complementary constraint. Kim argues that salaried incentive structures prevent the quality and productivity gains Tunguz measured from translating into organizational output volume. Only 5% of knowledge workers can use AI for production-quality output at all. Of those who can, salaried workers deliver existing output targets faster and stop. The gains disappear into leisure time, not additional output.
"Most people aren't incentivized to get more output. Look at incentivized teams — small teams, solo founders — they're showing how crazy good the output can get."
Kim's three-part throttle: skill gap, workflow defense, and misaligned incentives. When engineers manually review every AI change, the AI runs at human speed, not AI speed. The organizational structure caps the tool's output potential. AI productivity gains in salaried organizations cap at the existing output level because salary structures reward delivery, not volume.
- Tom Tunguz, AI raises the quality floor more than the ceiling by concentrating human effort on rhetorical refinement. Across 200 pieces, AI raised the 10th-percentile quality score by 1.22 points while keeping line-level editing constant, showing floor improvement outpaces ceiling improvement.
- Joseph Kim, AI productivity gains in salaried organizations cap at the existing output level because salary structures reward delivery, not volume. Salaried incentive structures, workflow defense, and a 5% skill gap combine to throttle AI productivity gains to zero marginal output in most knowledge organizations.
Game discovery through the citation graph
Joseph Kim and Kalie Moore audited three major LLMs on game recommendation prompts and found a consistent pattern: AI systems recommend games by citation authority in their training text, not by app store rank. Monument Valley, ranked #11 in the paid App Store, was the top recommendation from ChatGPT, Gemini, and Perplexity.
The structural reason: LLMs are trained on written text. Press coverage, buyer guides, and Wikipedia entries are in that text. Download counts and revenue rankings are not. Free-to-play games dominate charts but generate less written analysis than paid games because journalists and reviewers do not evaluate them the same way. The citation footprint of a paid game is larger per unit of popularity. Moore's audit finding makes this concrete:
"AI names games that are getting press coverage, and those are largely paid games."
HiRise earned Forbes and Business Insider coverage and outranked Roblox on 10 of 12 AI recommendation prompts. Across their sample, 74% of users chose the AI's top pick and 88% accepted the AI shortlist without additional searches. Kim's constraint on execution: "Do not use AI to write thought leadership. You lose everyone." LLMs recommend games by citation authority in training text, so studios without earned media coverage are invisible to AI-assisted discovery.
- Joseph Kim, LLMs recommend games by citation authority in training text, so studios without earned media coverage are invisible to AI-assisted discovery. AI recommendation systems draw on citation authority in training text, not download rank, making earned media the primary distribution lever for AI-assisted game discovery.
System prompts as reactive policy layers
Simon Willison documented a behavioral restriction appearing in Claude's published system prompt within days of Sony Music and Warner Chappell filing copyright lawsuits: a block on generating song lyrics. The pattern is one instance of a broader governance structure. System prompt edits take effect immediately, without model retraining, making them the fast layer of AI behavioral governance.
Willison's investigation also found that the published system prompt represents only the "core" system message. Feature-specific behavioral blocks, including end_conversation guidance, exist in separate unpublished layers. Published prompts are therefore incomplete transparency signals. Practitioners auditing model behavior from published documentation will consistently find undocumented constraints. Published AI system prompts are reactive legal compliance tools, not complete transparency signals, because unpublished feature-specific behavioral blocks remain hidden.
- Simon Willison, Published AI system prompts are reactive legal compliance tools, not complete transparency signals, because unpublished feature-specific behavioral blocks remain hidden. Published AI system prompts function as rapid-response legal compliance tools but understate actual behavioral constraints because feature-specific behavioral blocks remain unpublished.